Caitie Jones, Mallory Mills, Chay Newman and Max O’Shea

Are there appropriate parameters in place to report on the use of AI and therefore the environmental impacts of AI in Australian book publishing?

Keyterms: artificial intelligence (AI), Australian publishing industry, carbon emissions (CO2e), carbon footprint, greenwashing, infrastructural cost, Penguin Random House (PRH), transparent reporting

Introduction

Artificial intelligence (AI) and its benefits, uses and risks have risen to the forefront of conversation in various industries, the Australian publishing industry being no exception.

The rapid development of this technology, however, complicates these conversations, and the extent of its possible applications in publishing are yet to be seen. An even newer conversation pertains to the environmental sustainability of the often resource-intensive deployments of AI models—deployment referring to the act of taking a trained AI model and integrating it into a production environment. The emergent nature of the technology and the potential for profit means that the growth of AI, specifically generative AI, has ‘notably outpaced global regulatory efforts,’ resulting in ‘insufficient oversight of its socioeconomic and environmental impact’.

For the Australian publishing industry to fully understand how publishing practices impact the environment and how to make those practices more sustainable, the industry must be prepared for the impacts of AI use.

Initial research into AI

Emergent research into AI focuses primarily on the social and economic implications of AI use and the opportunities presented by certain AI models. Few studies are focused on the environmental impacts of AI and even fewer on the environmental impact in an industry-specific context.

An initial assessment of AI’s environmental footprint is offered by the Organisation for Economic Co-operation and Development (OECD). Their report outlines a definition of 'AI compute' and analyses the existing data and frameworks for environmental impact of AI computing, differentiating between direct and indirect environmental impact. While not publishing specific, these findings highlight the infrastructural cost of AI systems that the publishing industry is starting to depend on.

The OECD report makes several recommendations regarding transparency of reporting and measurement standards and recognises the difficulty of narrowing the focus from general computing to AI-specific data. It also notes the importance of broadening considerations of environmental impact from only energy consumption and carbon footprint—which has been the standard—to include other impacts, like water consumption and rare earth mining.

Another review of existing scientific literature on the environmental impacts of AI outlines the concept of ‘Green AI’, a framework for creating AI models in a way that reduces the computational resources required. While the framework is theoretical and makes recommendations primarily to developers, it demonstrates a way in which publishing houses can remain informed as consumers.

Other research into AI, however, notes the risks of unregulated growth and suggests that a narrow focus on improving energy efficiency, while important, is insufficient to address the impacts of AI compute and can further exacerbate the problem by leading to an overall increase in consumption. One study focusing on the use of AI in the book publishing industry in Ukraine observed that ‘one of the biggest challenges... is that the technology develops faster than people can perceive,’ so research into the impacts of AI development is often lagging behind.

Our research aims to identify the opportunities for the Australian book publishing industry to develop industry-specific measures to minimise the environmental footprint of AI. However, due to the emergent nature of the technology and the lack of relevant industry studies, there is significant need for further research.

Different types of AI and their current use in the publishing industry

The types of AI

AI is categorised by its capabilities, which are structured in an almost tier-like system: Narrow AI, General AI and Super AI. Narrow AI, or Weak AI, is trained to perform a specific function or task and therefore cannot think independently or overcome unforeseen challenges. This is the current state of AI.

General AI and Super AI exist only as concepts. General AI is defined by awareness, independence and consciousness; it can learn new skills, adapt in real-time and solve complex medical issues. It sounds impossible, but one 2017 report notes that this was the ‘original goal’ before AI research turned more problem specific.

Super AI is an interface so advanced it far surpasses the capabilities of humanity in every conceivable area. It currently remains a fantasy, present only within the fictitious parameters of television and film.

AI in publishing

Narrow AI is the only form of artificial intelligence currently being utilised by the publishing industry—and compared to something like Super AI, its capabilities appear quite limited. But Narrow AI has already permeated many areas of the publishing industry. One study by communications expert, Oleksii Sytnyk, analysed the successful implementations of AI within five different publishing houses.

In 2019, German publishing house Springer Nature published the first scientific book partially generated by AI. Throughout the book’s creation, AI was used to analyse articles and summarise the results, which allowed the work to be completed much quicker—although humans still performed quality and accuracy checks.

In 2022, Penguin Random House introduced a system that analysed manuscripts and predicted potential bestsellers, while Elsevier used AI to prepare and review scientific publications to identify plagiarism. Authors at Elsevier were advised to only use AI to improve readability and language, and not to replace conclusions, key advice or medical ideas; as such, Elsevier claimed authors were fully responsible for the content of their work.

Other ways these five publishing houses used AI included conducting research, finding reviewers in specific research fields, and making editorial suggestions—generally time-consuming and labour-intensive tasks. Work that would ‘otherwise take hours’ can now be completed in less than a minute.

Sytnyk concluded their study by claiming that ‘technophobia’ and the conservative nature of the publishing industry were responsible for much of the resistance around AI. They credited AI with ‘increased efficiency, improved content quality and optimization of editorial processes.’ And while they noted AI should be used in tandem with human employees as a tool, not a replacement, they suggested publishing houses would lose their ‘competitiveness’ if they failed to ‘innovate.’

Another study detailed an algorithm used by European e-book and audiobook distributor Bookwire, which ‘match[es] and compar[es] various historical data points for prices and track[s] the title performance throughout time.’ It generates suggested ideal prices and limited special promotions. But like Sytnyk, the authors emphasised that AI is a tool and that human staff members are essential in successfully utilising it. ‘AI is very specific in the tasks it carries out and isn’t universally applicable […] It is not the human mind.’

Dr Dang Nguyen, a research fellow at RMIT specialising in automated decision making, disagrees. ‘Framing AI’s “presence” in terms of advantages and disadvantages is reductive,’ said Nguyen in an online interview. ‘It treats AI as a neutral tool rather than a sociotechnical system embedded in power, labour and infrastructure.’ Often, Nguyen explains, AI isn’t developed with a specific need in mind, rather, it begins with ‘hype cycles, funding opportunities and the search for a competitive edge.’ This ‘competitive edge’ relies on the idea of Super AI, which proposes innovation so advanced it supersedes the human brain.

The natural resources consumed by AI in its current form are so immense that ‘a common AI training model in Linguistics can emit more than 284 tonnes [284,000 kg] of carbon dioxide equivalent.’ So the idea of an inhuman super intelligence capable of replacing our jobs—and exhausting our environment—naturally leaves many people hesitant.

Narrow AI might only specialise in specific tasks, but its existence suggests the possibility of greater advancements to come.

For now, the consideration is whether the environmental costs are worth reducing a few hours of labour. AI is often touted as an environmental solution, but Nguyen warns against investing in ‘the mythology of “problem-solving tech”.’ Even if AI is used to track and analyse environmental data, the cost of this technology possibly outweighs the potential benefits. ‘The question shouldn’t be “is AI being developed to solve problems?”’ says Nguyen, ‘but: what counts as a problem worth solving, who defines it and who benefits from the so-called solution?’

Risks of profit-driven AI

One of the few comprehensive resources available on the use of AI in the publishing industry is a 2020 report produced by Frontier Economics on behalf of the Publishers Association in the United Kingdom. There are several indications from the benefits and barriers outlined in this report that the Green AI framework—AI models designed for ‘efficiency and reduced environmental impact’—may not be considered by publishing houses when implementing AI technology.

Publishing houses already using AI technology have commonly cited ‘increased competitive advantage’ as a current and potential benefit. These companies tend to prioritise the accuracy, speed and performance of the models they develop and use. As a result, newly developed AI models are more likely to emphasise profit and fit into the category of ‘Red’ AI.

Additionally, many publishing houses cited a lack of available training data as a potential barrier for AI, with some investing ‘considerable resources […] to add metadata to their back catalogue’ to satisfy the ‘increasingly large volumes of data [required by AI] to train algorithms.’ This indicates that AI development focuses on maximising the resources available to increase performance, rather than minimising resources needed, as Green AI tries to do.

Though the report focuses on investment into the AI sector, it makes no reference to the ‘well-known diminishing returns of increased [computational] cost’ in training AI, which suggests a point at which the additional resources used by a training program require ‘an exponentially larger model’ to yield a ‘linear gain in performance’. This applies to increases in training data and experiments, not only computational power. For example, the cost of developing an AI model is proportional to the cost of processing a single example, times the size of the training dataset and the number of experiments.

[Insert AI_Info 1]

One of the most contentious applications of AI in the publishing industry is the idea of training a model to perform creative content creation, like writing a novel. Training and running this kind of model would, realistically, require significant computational power, enormous amounts of training material (likely the copyrighted work of human authors) and numerous experiments to process even a single example.

The sustainability reporting framework generally used by corporations, like Penguin Random House’s Our Sustainability Story reports, focuses on measures such as carbon emissions, which some researchers posit is an impractical measure for AI models. It also focuses on reporting based on three categories based on the source of the emissions: Scopes 1 (direct sources like onsite generators), 2 (indirect sources like imported electricity consumption) and 3 (upstream and downstream sources).

To accurately account for AI model development and deployment, publishing houses need to understand how AI falls under these parameters and how they should account for impacts beyond these metrics. In her keynote speech ‘The Ecological Impact of an Automated Society,’ delivered at the 2022 ADM+S Symposium, Dr Melissa Gregg illustrates that the risk of this reporting could be that corporations reframe efforts to minimise their environmental impacts, to instead minimise their sustainability reports.

Reporting environmental impacts of AI from a third-party company source

The inclusion of AI’s environmental costs in sustainability reports is a relatively new issue. One major problem is that most third-party AI providers do not give detailed enough descriptions of their products. The OECD argues that AI manufacturers generally report energy consumption but do not account for the energy used for the manufacture, transportation or disposal of AI hardware. Lack of data on indirect emissions means that companies that use AI are left with an incomplete idea of its actual environmental impact. This negatively impacts publishing houses who wish to accurately report indirect emissions.

Nguyen notes that companies may know about AI’s environmental impact, such as energy use and e-waste, but that ‘corporate narratives often focus on surface-level sustainability metrics.’ They focus on the use of renewable energy but turn a blind eye to ‘deeper political and infrastructural consequences of datafication’, creating a ‘depoliticised and compartmentalised’ vision of responsibility for the environment.

Furthermore, there are risks like greenwashing, where companies exaggerate their sustainability progress, whether deliberately or through incomplete reporting. For example, if a publishing house openly communicates about the use of AI in their work but does not say anything about the energy consumption, it may be viewed as greenwashing. To address this, the industry should encourage the standardisation of environmental disclosures by AI companies so that sustainability reports include a proper estimation of AI related emissions.

So far, most reporting on environmental sustainability in the publishing industry does not acknowledge the negative environmental effects of AI, partly because many industries do not have reporting standards that properly capture AI’s resource requirements. It is also worth noting that current legislation does not require that level of reporting. In addition, publishing houses may lack the capabilities and resources to report on these impacts.

A review on greenwashing discovered that although some AI consumption data is disclosed, companies do not report other environmental metrics like the amount of water that data centres use or the greenhouse gas (GHG) emissions involved in manufacturing AI hardware.

As Nguyen says, the publishing industry should report all third-party sustainability data, including labour exploitation and resource extraction. This would shift the focus from ‘PR friendly metrics to structural accountability’ across the AI life cycle.

To address this gap, publishing houses can incorporate third-party AI data into their sustainability reports.

Some of the challenges of determining the third-party carbon footprint include the absence of standard formats for measuring AI’s environmental impact. Current sustainability frameworks, including the Global Reporting Initiative Standards and the Sustainability Accounting Standard Board Standards, do not incorporate AI-related specifications. Consequently, AI specific emissions are sometimes included under general information and communication technology emissions, which may hide their unique environmental impact.

In response, the publishing industry must campaign for the adoption of AI-focused reporting standards. Such policies should cover the entire AI life cycle, from the sourcing of raw materials for manufacturing the hardware, to the disposal of electronic waste from the systems.

For instance, the publishing industry could demand that AI model training includes disclosure on the carbon footprint. Supply chain transparency should also be initiated to check the sustainability of claims made by AI companies.

Mo Fanning, an author and publisher, argues that overlooking AI’s sustainability impact may negate other advances in the publishing industry, like the switch to recycled paper and print on demand. This illustrates the need for proper parameters to prevent the introduction of AI from causing a regression on environmental sustainability efforts. Nguyen recommends that ‘companies shouldn’t just look for “sustainable practices”—they should interrogate the unsustainable ones they already depend on.’

Currently, the publishing industry sources sustainability reporting data through internal audits and from their suppliers’ voluntary disclosures. However, this approach is insufficient to account for the environmental impact of AI, since many third-party providers do not report detailed emissions data.

Despite AI’s increasing use, most companies have failed to design mechanisms to monitor the environmental impact of AI. This is in line with Nguyen's argument that AI is developing at such a rapid pace that some companies are more focused on ‘technical efficiency or security’ than ethics, disregarding the full environmental impact and infrastructure inequities.

The publishing industry needs to address the greenwashing issue in sustainability reporting, where organisations present an environmental record while leaving out important facts. When reporting sustainability efforts, the publishing industry needs to be aware of unintentionally hiding data, infrastructural disparities and environmental costs associated with AI.

Measuring the cost/benefit of AI

Costs of AI

AI systems and their effects on the global environment are difficult to quantify. There are a couple of things to contend with when considering AI in a corporate context. Firstly, there is no ‘universally accepted definition’ of AI. Secondly, the United Nations Environment Programme (UNEP) Issue Notice highlights both the lack of comprehensive research into AI and standardisation of reporting methods on the environmental impact of AI.

The OECD notes that assessments of AI’s environmental impact largely focus on energy use and GHG emissions. Coal, the most carbon-intensive fossil fuel, still supplies just over a third of global electricity generation, while oil and natural gas account for 15% of global energy sector GHG emissions.

[Insert AI_Info2]

Their report advocates for expanding this exploration to impacts such as water consumption, waste management and biodiversity.

The AI life cycle can be considered through two stages: software and hardware. These stages have different considerations for environmental sustainability. The entire life cycle of AI must be considered when thinking about its environmental costs.

[Insert AI_Info3]

The use of freshwater in every stage of AI is accelerating. It could be over 6.6 billion cubic metres (m³) by 2027, which is troubling considering that over 1.1 billion people currently experience water scarcity. Gregg states that ‘water is critical to computer chip production from start to finish,’ functioning as a coolant for data centres and a cleaner for metals and chemical compounds.

The equipment used to train and run AI models also produces e-waste, consisting of electrical and electronic equipment that cannot be disposed of with regular waste. Only 22% of e-waste is being formally collected and recycled today, according to the 2024 Global E-Waste Monitor. AI could generate 1.2 to 5 million tonnes of e-waste by 2030.

The negative effects of AI on the environment could be seen as insurmountable, considering the current state of AI and its future use.

Benefits of AI

The future impact of artifical intelligence on the publishing industry, a joint study between book fair Frankfurter Buchmesse and management consultancy Gould Finch, contextualises the benefits of AI in three key findings:

1) Artificial intelligence is not going to replace writers, but it is able to strengthen core-business.

AI programs, while able to apply basic proofreading, collating, prose and tone skills in writing, have no capability to create quality pieces of writing. Core business procedures such as marketing, production, administration and analytics are targets of AI programs.

2) Minimal investments can still bring in monetary benefits.

Investing even minimally in AI programs or employee training in AI can lead to monetary benefits which outweigh the expenditure.

3) Investing in artificial intelligence does not mean fewer jobs for humans.

Companies who have implemented AI are finding greater job security and positive effects on reader statistics and overall sales. The World Economic Forum expands on this in Creative Disruption: The impact of emerging technologies on the creative economy:

‘In particular, AI that generates text is widespread in journalism and is used by publishers to expand the range of offerings. The Associated Press has used AI to free up around 20% of reporters’ time while increasing output tenfold. The Washington Post developed its own tool, Heliograf, to cover sports and political news. In its first year it generated about 70 articles a month, mostly stories it would not have dedicated staff to.’

While the use of AI can result in monetary benefits through ‘clicks’, it still does not seem justified when considering the extensive environmental costs. Nguyen says: ‘while there are promising applications, they often rest on infrastructures that raise serious environmental concerns.’

Findings

Gregg references the risks of ‘the language of projection,’ which focuses on the impending impacts of a technology but removes accountability for the present impact of those decisions. She talks about how the ongoing discussion regarding the future environmental impacts of computing often glosses over the discussion of the impacts that are already clearly visible.

The conversation around sustainability reporting needs to be reframed similarly to understand the impacts of AI. This has been linked not only to carbon emissions, but also resource depletion and a risk of increased socio-economic disparity.

This topic needs, and deserves, more (quality) research. There is huge potential for the Australian publishing industry to collaborate in research with other related industries, such as media and journalism, to understand the effects of developing new AI models and deploying AI. This research should be done by openly reporting on and taking accountability for current impacts.

Frontier Economics suggested that publishing houses adopt a ‘benefit-led approach to their AI investment decisions.’ However, using a cost–benefit analysis on such an emergent technology with so many variables can be incredibly difficult. As Nguyen said, ‘AI development doesn’t begin with a real, grounded need.’

Conclusion

While the publishing industry has adopted Narrow AI technologies to summarise key points in research or to track title performance over time, the emergent nature of the technology and the lack of research has consequences. It cannot be universally reported across the industry as there is no standard. Its environmental costs cannot be quantified as there is no hard data and, as such, it is still largely speculation.

The emergent nature of the technology, however, also offers an opportunity. As the publishing industry begins to develop policies and codes of conduct around AI use, we can foreground questions of environmental sustainability into these policies from the outset, so that they become an integral consideration rather than an afterthought.